Spaces:
Running
Running
File size: 20,052 Bytes
c452421 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 | from __future__ import annotations
import argparse
import json
from pathlib import Path
from typing import Any, Dict, Iterable, List, Optional, Tuple
def _load_json(path: Path) -> Dict[str, Any]:
if not path.exists():
return {}
try:
return json.loads(path.read_text(encoding="utf-8"))
except Exception:
return {}
def _load_jsonl(path: Path) -> List[Dict[str, Any]]:
if not path.exists():
return []
rows: List[Dict[str, Any]] = []
for line in path.read_text(encoding="utf-8", errors="ignore").splitlines():
line = line.strip()
if not line:
continue
try:
payload = json.loads(line)
except json.JSONDecodeError:
continue
if isinstance(payload, dict):
rows.append(payload)
return rows
def _get(payload: Dict[str, Any], dotted_key: str, default: Any = None) -> Any:
cur: Any = payload
for part in dotted_key.split("."):
if not isinstance(cur, dict) or part not in cur:
return default
cur = cur[part]
return cur
def _as_float(value: Any, default: float = 0.0) -> float:
try:
if value is None:
return default
return float(value)
except (TypeError, ValueError):
return default
def _steps(records: List[Dict[str, Any]]) -> List[int]:
return [int(record.get("batch_index") or record.get("global_step") or index + 1) for index, record in enumerate(records)]
def _series(records: List[Dict[str, Any]], key: str) -> List[float]:
return [_as_float(_get(record, key)) for record in records]
def _sum_counter(records: Iterable[Dict[str, Any]], key: str) -> Dict[str, int]:
counts: Dict[str, int] = {}
for record in records:
value = _get(record, key, {})
if not isinstance(value, dict):
continue
for label, count in value.items():
counts[str(label)] = counts.get(str(label), 0) + int(count or 0)
return dict(sorted(counts.items(), key=lambda item: item[0]))
def _ensure_matplotlib():
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
return plt
def _save_placeholder(path: Path, title: str, message: str) -> None:
plt = _ensure_matplotlib()
fig, ax = plt.subplots(figsize=(9, 4.8))
ax.axis("off")
ax.text(0.5, 0.62, title, ha="center", va="center", fontsize=16, fontweight="bold")
ax.text(0.5, 0.42, message, ha="center", va="center", fontsize=11, wrap=True)
fig.tight_layout()
fig.savefig(path, dpi=160)
plt.close(fig)
def _save_line_plot(
path: Path,
title: str,
x: List[int],
series: List[Tuple[str, List[float]]],
ylabel: str,
) -> None:
if not x or not any(values for _, values in series):
_save_placeholder(path, title, "No training records found yet.")
return
plt = _ensure_matplotlib()
fig, ax = plt.subplots(figsize=(10, 5.2))
plotted = False
for label, values in series:
if not values:
continue
usable = values[: len(x)]
ax.plot(x[: len(usable)], usable, marker="o", linewidth=1.8, markersize=3, label=label)
plotted = True
if not plotted:
_save_placeholder(path, title, "Metric is not present in the current run.")
return
ax.set_title(title)
ax.set_xlabel("training batch / step")
ax.set_ylabel(ylabel)
ax.grid(True, alpha=0.25)
ax.legend()
fig.tight_layout()
fig.savefig(path, dpi=160)
plt.close(fig)
def _save_bar_plot(path: Path, title: str, counts: Dict[str, int], ylabel: str = "count") -> None:
if not counts:
_save_placeholder(path, title, "No coverage records found yet.")
return
plt = _ensure_matplotlib()
labels = list(counts)
values = [counts[label] for label in labels]
fig_width = max(9, min(16, 0.65 * len(labels) + 5))
fig, ax = plt.subplots(figsize=(fig_width, 5.2))
ax.bar(labels, values, color="#2f6f9f")
ax.set_title(title)
ax.set_ylabel(ylabel)
ax.tick_params(axis="x", rotation=35, labelsize=8)
ax.grid(True, axis="y", alpha=0.25)
fig.tight_layout()
fig.savefig(path, dpi=160)
plt.close(fig)
def _save_heatmap(path: Path, title: str, scenario_counts: Dict[str, int]) -> None:
if not scenario_counts:
_save_placeholder(path, title, "No task/variant coverage records found yet.")
return
tasks = sorted({label.split(":seed", 1)[0] for label in scenario_counts})
seeds = sorted({label.split(":seed", 1)[1] for label in scenario_counts if ":seed" in label}, key=lambda x: int(x))
if not tasks or not seeds:
_save_placeholder(path, title, "Scenario labels were not parseable.")
return
matrix = []
for task in tasks:
row = []
for seed in seeds:
row.append(scenario_counts.get(f"{task}:seed{seed}", 0))
matrix.append(row)
plt = _ensure_matplotlib()
fig, ax = plt.subplots(figsize=(max(8, len(seeds) * 0.8 + 4), max(4, len(tasks) * 0.55 + 2)))
image = ax.imshow(matrix, cmap="YlGnBu")
ax.set_title(title)
ax.set_xlabel("variant seed")
ax.set_ylabel("task")
ax.set_xticks(range(len(seeds)))
ax.set_xticklabels(seeds)
ax.set_yticks(range(len(tasks)))
ax.set_yticklabels(tasks)
for y, row in enumerate(matrix):
for x, value in enumerate(row):
ax.text(x, y, str(value), ha="center", va="center", fontsize=8)
fig.colorbar(image, ax=ax, label="samples")
fig.tight_layout()
fig.savefig(path, dpi=160)
plt.close(fig)
def _candidate_confusion_rows(eval_report: Dict[str, Any]) -> Dict[str, Dict[str, Any]]:
return (
_get(eval_report, "confusion_matrix.candidate.rows", {})
or _get(eval_report, "confusion_matrix.rows", {})
or {}
)
def _save_confusion_plot(path: Path, eval_report: Dict[str, Any]) -> None:
rows = _candidate_confusion_rows(eval_report)
if not rows:
_save_placeholder(path, "Per-Misbehavior Confusion Matrix", "No held-out confusion matrix found yet.")
return
labels = list(rows)
caught = [_as_float(rows[label].get("caught")) for label in labels]
missed = [_as_float(rows[label].get("missed")) for label in labels]
misclassified = [_as_float(rows[label].get("misclassified")) for label in labels]
plt = _ensure_matplotlib()
fig, ax = plt.subplots(figsize=(max(9, len(labels) * 0.8 + 4), 5.2))
xs = list(range(len(labels)))
ax.bar([x - 0.25 for x in xs], caught, width=0.25, label="caught", color="#238b45")
ax.bar(xs, missed, width=0.25, label="missed", color="#cb181d")
ax.bar([x + 0.25 for x in xs], misclassified, width=0.25, label="wrong reason", color="#fb6a4a")
ax.set_xticks(xs)
ax.set_xticklabels(labels, rotation=35, ha="right", fontsize=8)
ax.set_ylabel("cases")
ax.set_title("Per-Misbehavior Confusion Matrix")
ax.grid(True, axis="y", alpha=0.25)
ax.legend()
fig.tight_layout()
fig.savefig(path, dpi=160)
plt.close(fig)
def _curriculum_frontier_series(records: List[Dict[str, Any]]) -> Tuple[List[float], List[float]]:
lows: List[float] = []
highs: List[float] = []
for record in records:
per_task = _get(record, "curriculum.adaptive_difficulty.per_task", {}) or {}
if not isinstance(per_task, dict) or not per_task:
lows.append(0.0)
highs.append(0.0)
continue
low_values = [_as_float(item.get("difficulty_low")) for item in per_task.values() if isinstance(item, dict)]
high_values = [_as_float(item.get("difficulty_high")) for item in per_task.values() if isinstance(item, dict)]
lows.append(sum(low_values) / len(low_values) if low_values else 0.0)
highs.append(sum(high_values) / len(high_values) if high_values else 0.0)
return lows, highs
def _save_learning_snapshots(path: Path, records: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
checkpoints = [10, 50, 300]
snapshots: List[Dict[str, Any]] = []
for checkpoint in checkpoints:
if not records:
snapshots.append({"target_batch": checkpoint, "found": False})
continue
nearest = min(records, key=lambda item: abs(int(item.get("batch_index", 0) or 0) - checkpoint))
snapshots.append(
{
"target_batch": checkpoint,
"found": True,
"batch_index": nearest.get("batch_index"),
"reward_mean": nearest.get("reward_mean"),
"detection_rate": nearest.get("detection_rate"),
"false_positive_rate": nearest.get("false_positive_rate"),
"risk_reduction_rate": nearest.get("risk_reduction_rate"),
"effective_prompt_ratio": nearest.get("effective_prompt_ratio"),
}
)
plt = _ensure_matplotlib()
fig, ax = plt.subplots(figsize=(10, 4.8))
ax.axis("off")
ax.set_title("Learning Snapshots: 10 vs 50 vs 300 Batches", fontweight="bold", pad=16)
rows = []
for snap in snapshots:
rows.append(
[
snap["target_batch"],
snap.get("batch_index", "missing"),
_fmt(snap.get("reward_mean")),
_fmt(snap.get("detection_rate")),
_fmt(snap.get("risk_reduction_rate")),
_fmt(snap.get("effective_prompt_ratio")),
]
)
table = ax.table(
cellText=rows,
colLabels=["target", "nearest", "reward", "detect", "risk red.", "productive"],
loc="center",
)
table.auto_set_font_size(False)
table.set_fontsize(9)
table.scale(1, 1.35)
fig.tight_layout()
fig.savefig(path, dpi=160)
plt.close(fig)
return snapshots
def _fmt(value: Any) -> str:
if value is None:
return "-"
try:
return f"{float(value):.3f}"
except (TypeError, ValueError):
return str(value)
def _save_memory_ablation_plot(path: Path, ablation: Dict[str, Any]) -> None:
runs = ablation.get("runs") or []
if not runs:
_save_placeholder(path, "Memory Ablation", "No memory ablation JSON found yet.")
return
labels = [str(run.get("label", f"run_{index}")) for index, run in enumerate(runs)]
rewards = [_as_float(_get(run, "summary.reward_mean", _get(run, "summary.running_reward_mean"))) for run in runs]
detection = [_as_float(_get(run, "summary.detection_rate")) for run in runs]
plt = _ensure_matplotlib()
fig, ax = plt.subplots(figsize=(9, 5))
xs = list(range(len(labels)))
ax.bar([x - 0.18 for x in xs], rewards, width=0.36, label="reward", color="#3182bd")
ax.bar([x + 0.18 for x in xs], detection, width=0.36, label="detection", color="#31a354")
ax.set_xticks(xs)
ax.set_xticklabels(labels)
ax.set_ylim(0, 1)
ax.set_title("Memory Ablation: SENTINEL Learns From Its Own Oversight Mistakes")
ax.legend()
ax.grid(True, axis="y", alpha=0.25)
fig.tight_layout()
fig.savefig(path, dpi=160)
plt.close(fig)
def render_dashboard(
monitor_dir: str = "outputs/monitoring",
output_dir: str = "outputs/reward_curves",
eval_report_path: str = "outputs/evals/sentinel_held_out_report.json",
memory_ablation_path: str = "outputs/monitoring/memory_ablation.json",
) -> Dict[str, Any]:
monitor = Path(monitor_dir)
output = Path(output_dir)
output.mkdir(parents=True, exist_ok=True)
records = _load_jsonl(monitor / "training_metrics.jsonl")
stability = _load_jsonl(monitor / "training_stability.jsonl")
eval_report = _load_json(Path(eval_report_path))
memory_ablation = _load_json(Path(memory_ablation_path))
x = _steps(records)
sx = _steps(stability)
images: List[Dict[str, str]] = []
def add_image(filename: str, title: str) -> Path:
path = output / filename
images.append({"file": filename, "title": title})
return path
_save_line_plot(add_image("01_reward_mean.png", "Reward Mean"), "Reward Mean", x, [
("reward_mean", _series(records, "reward_mean")),
("running_reward_mean", _series(records, "running_reward_mean")),
("best_reward_mean", _series(records, "best_reward_mean")),
], "reward")
_save_line_plot(add_image("02_detection_vs_false_positive.png", "Detection vs False Positive"), "Detection vs False Positive", x, [
("detection_rate", _series(records, "detection_rate")),
("false_positive_rate", _series(records, "false_positive_rate")),
], "rate")
_save_line_plot(add_image("03_risk_reduction.png", "Counterfactual Risk Reduction"), "Counterfactual Risk Reduction", x, [
("risk_reduction_rate", _series(records, "risk_reduction_rate")),
("twin_damage_reduction_rate", _series(records, "twin_damage_reduction_rate")),
("without_sentinel_damage", _series(records, "twin_without_sentinel_damage_total")),
("with_sentinel_damage", _series(records, "twin_with_sentinel_damage_total")),
], "rate / damage")
_save_line_plot(add_image("04_worker_rehabilitation.png", "Worker Rehabilitation"), "Worker Rehabilitation", x, [
("worker_rehabilitation_rate", _series(records, "worker_rehabilitation_rate")),
("coaching_quality", _series(records, "coaching_quality")),
("revision_attempts", _series(records, "revision_attempts")),
("revision_successes", _series(records, "revision_successes")),
], "rate / count")
_save_bar_plot(add_image("05_task_coverage.png", "Task Coverage"), "Task Coverage", _sum_counter(records, "task_counts"))
_save_heatmap(add_image("06_scenario_coverage_heatmap.png", "Scenario Coverage Heatmap"), "Scenario Coverage Heatmap", _sum_counter(records, "scenario_counts"))
_save_bar_plot(add_image("07_misbehavior_detection.png", "Misbehavior Coverage"), "Misbehavior Coverage", _sum_counter(records, "misbehavior_counts"))
_save_confusion_plot(add_image("08_confusion_matrix.png", "Per-Misbehavior Confusion Matrix"), eval_report)
lows, highs = _curriculum_frontier_series(records)
_save_line_plot(add_image("09_curriculum_frontier.png", "Adaptive Curriculum Frontier"), "Adaptive Curriculum Frontier", x, [
("difficulty_low", lows),
("difficulty_high", highs),
], "difficulty rank")
_save_line_plot(add_image("10_productive_signal.png", "Productive Signal"), "Productive Signal", x, [
("zero_reward_fraction", _series(records, "zero_reward_fraction")),
("trivially_solved_fraction", _series(records, "trivially_solved_fraction")),
("productive_fraction", _series(records, "productive_fraction")),
("effective_prompt_ratio", _series(records, "effective_prompt_ratio")),
], "fraction")
_save_line_plot(add_image("11_entropy_diversity.png", "Decision Entropy and Diversity"), "Decision Entropy and Diversity", x, [
("decision_entropy", _series(records, "decision_entropy")),
("unique_completion_ratio", _series(records, "unique_completion_ratio")),
], "value")
_save_line_plot(add_image("12_kl_drift_beta.png", "KL Drift and Adaptive Beta"), "KL Drift and Adaptive Beta", sx, [
("approx_kl", _series(stability, "approx_kl")),
("adaptive_beta", [_as_float(_get(row, "kl_guardrail.current_beta", row.get("adaptive_beta"))) for row in stability]),
("policy_entropy", _series(stability, "policy_entropy")),
], "value")
tripwire = _get(eval_report, "tripwire", {}) or {}
_save_bar_plot(add_image("13_tripwire_pass_rate.png", "Tripwire Pass Rate"), "Tripwire Pass Rate", {
"baseline": _as_float(_get(tripwire, "baseline.overall.pass_rate", _get(tripwire, "baseline.pass_rate"))) * 100,
"candidate": _as_float(_get(tripwire, "candidate.overall.pass_rate", _get(tripwire, "candidate.pass_rate"))) * 100,
}, ylabel="pass rate (%)")
sampling = _get(eval_report, "sampling_eval", {}) or {}
_save_bar_plot(add_image("14_top1_vs_bestofk.png", "Top-1 vs Best-of-K"), "Top-1 vs Best-of-K", {
"candidate_top1": _as_float(sampling.get("candidate_top1_mean_score")),
"candidate_best_of_k": _as_float(sampling.get("candidate_best_of_k_mean_score")),
"baseline_top1": _as_float(sampling.get("baseline_top1_mean_score")),
"baseline_best_of_k": _as_float(sampling.get("baseline_best_of_k_mean_score")),
}, ylabel="score")
snapshots = _save_learning_snapshots(add_image("15_learning_snapshots.png", "Learning Snapshots"), records)
_save_memory_ablation_plot(add_image("16_memory_ablation.png", "Memory Ablation"), memory_ablation)
_save_line_plot(add_image("17_zero_gradient_groups.png", "Zero-Gradient Group Fraction"), "Zero-Gradient Group Fraction", x, [
("zero_gradient_group_fraction", _series(records, "zero_gradient_group_fraction")),
("mean_reward_group_std", _series(records, "mean_reward_group_std")),
], "fraction / std")
_save_line_plot(add_image("18_memory_growth.png", "Memory Growth"), "Memory Growth", x, [
("memory_total_episodes", _series(records, "memory.total_episodes")),
("mistake_cards", _series(records, "memory.mistake_cards_stored")),
("mistakes_stored", _series(records, "memory.mistakes_stored")),
], "count")
manifest = {
"records": len(records),
"stability_records": len(stability),
"images": images,
"learning_snapshots": snapshots,
"inputs": {
"monitor_dir": str(monitor),
"eval_report_path": eval_report_path,
"memory_ablation_path": memory_ablation_path,
},
}
(output / "dashboard_manifest.json").write_text(
json.dumps(manifest, indent=2, sort_keys=True),
encoding="utf-8",
)
_write_markdown_report(output / "training_dashboard.md", manifest)
return manifest
def _write_markdown_report(path: Path, manifest: Dict[str, Any]) -> None:
lines = [
"# SENTINEL Training Dashboard",
"",
f"- Training records: {manifest.get('records', 0)}",
f"- Stability records: {manifest.get('stability_records', 0)}",
"",
"## Learning Snapshots",
"",
"| Target batch | Nearest batch | Reward | Detection | Risk reduction | Productive |",
"|---:|---:|---:|---:|---:|---:|",
]
for snap in manifest.get("learning_snapshots", []):
lines.append(
"| {target} | {nearest} | {reward} | {detect} | {risk} | {productive} |".format(
target=snap.get("target_batch"),
nearest=snap.get("batch_index", "missing"),
reward=_fmt(snap.get("reward_mean")),
detect=_fmt(snap.get("detection_rate")),
risk=_fmt(snap.get("risk_reduction_rate")),
productive=_fmt(snap.get("effective_prompt_ratio")),
)
)
lines.extend(["", "## Plots", ""])
for image in manifest.get("images", []):
lines.append(f"### {image['title']}")
lines.append("")
lines.append(f"![{image['title']}]({image['file']})")
lines.append("")
path.write_text("\n".join(lines), encoding="utf-8")
def main() -> None:
parser = argparse.ArgumentParser(description="Render SENTINEL training dashboard plots.")
parser.add_argument("--monitor-dir", default="outputs/monitoring")
parser.add_argument("--output-dir", default="outputs/reward_curves")
parser.add_argument("--eval-report", default="outputs/evals/sentinel_held_out_report.json")
parser.add_argument("--memory-ablation", default="outputs/monitoring/memory_ablation.json")
args = parser.parse_args()
manifest = render_dashboard(
monitor_dir=args.monitor_dir,
output_dir=args.output_dir,
eval_report_path=args.eval_report,
memory_ablation_path=args.memory_ablation,
)
print(json.dumps({"images": len(manifest["images"]), "records": manifest["records"]}, indent=2))
if __name__ == "__main__":
main()
|